KR101038402B1 - Real-time capturing and generating stereo images and videos with a monoscopic low power mobile device - Google Patents

Abstract

Monoscopic low-power mobile devices can generate real-time stereo images and video from a single captured view. The device uses the statistics from the autofocusing process to generate a block depth map of a single captured view. Artifacts in the block depth map are reduced and an image depth map is generated. Using a disparity map and a Z-buffer based 3D surface reconstruction process that is a function of the geometry of the binocular vision, stereo three-dimensional (3D) left and right views are generated from the image depth map.

Monoscopic low-power devices, still images, video images

Description

REAL-TIME CAPTURING AND GENERATING STEREO IMAGES AND VIDEOS WITH A MONOSCOPIC LOW POWER MOBILE DEVICE}

Background of the Invention

Technical Field

In general, the present invention relates to a monoscopic low-power mobile device, such as a hand-held camera, a camcorder, a single-sensor camera phone, or another single camera sensor device capable of generating real-time images and videos. . The invention also relates to a real-time stereo image, a method of generating a still image capture device, and a video image capture device.

Background

In recent years, perceptual realism has become one of the major forces driving the transformation of the next generation of multimedia development. The rapidly evolving multimedia communications and entertainment market calls for 3D stereoscopic image and video technologies covering stereo image capture, processing, compression, delivery, and display. Several attempts at future standards such as 3DTV and MPEG 3 DAV have been launched to satisfy such requests.

The main difference between a stereo image and a mono image is that the stereo image provides a feeling of three-dimensional and distance to objects in the scene. Innate human vision is stereoscopic due to the binocular view observed by the left and right eyes with different perspective viewpoints. The human brain can synthesize images at stereoscopic depths. In general, a stereoscopic camera with two sensors is required to generate a stereoscopic image or video. However, most multimedia devices currently in use are implemented within a monoscopic infrastructure.

In the past years, stereoscopic image generation has been actively studied. In one study, a video sequence is analyzed and a 3D scene structure is estimated from 2D geometry and motion activity (also called structure from motion (SfM)). This class of approaches enables the conversion of recorded 2D video chips to 3D. However, since the computational complexity is rather high, it is not possible for real-time stereo image generation. On the other hand, since SfM is a mathematically ill-posed problem, the result may include artifacts and may cause visual inconvenience. First, some other approaches estimate depth information from a single-view still-image based on a heuristic rule according to a particular application, and then generate a stereoscopic view accordingly.

In another study, a method of extracting relative depth information from a monoscopic cue, eg, the retina size of an object, is proposed, which is useful for generating secondary depth maps. In another study, a facial feature based on a parametric depth map generation scheme is proposed for converting 2D head-and-shoulder images to 3D. In another proposed method for depth-map generation, some of the steps of this approach, for example image classification in preprocessing, are not trivial and the implementation may be very complex and the practicality of the proposed algorithm Weakens. In another method, a 2D image real-time conversion algorithm to 3D using motion detection and region segmentation is proposed. However, due to inaccuracies in object segmentation and object depth estimation, artifacts cannot be avoided. Clearly, all the above mentioned methods only consider captured monoscopic images. Some other approaches use ancillary sources to assist in generating stereo views. For example, a low cost secondary monochrome or low resolution camera is used to capture additional views, and then using a disparity estimation model to create a depth map of the pixels.

In another example, a monoscopic high resolution color camera is used to capture the brightness and chromaticity of a scene, and the cheap flanking 3-D stereoscopic pair of low resolution monochrome “outrigger” cameras has a depth of brightness. And to increase chromaticity. The disparity map generated from the three views obtained is used to synthesize stereoscopic pairs. In another example, a mixed set of automatic and manual techniques are used to extract the depth map (often an automatic method is not reliable), and then a simple smoothing filter is used to visualize the artifacts of the resulting image. Decrease.

As can be readily appreciated, there is a need for a low-complexity method for obtaining real-time stereo images and videos by using a monoscopic mobile camera phone or other low-power monoscopic device.

There is a need for monoscopic low-power mobile devices that estimate depth map information in a manner that avoids computational complexity induced by using SfM or depth analysis as well as auxiliary equipment or human-interaction used in other approaches. do.

There is also a need for a monoscopic low-power mobile device that uses a low-complexity approach to detect and estimate depth information for real-time capture and generation of stereo video.

Summary of the Invention

In view of the foregoing, it is an object of the present invention to provide a monoscopic low-power mobile device that utilizes a low-complexity approach for detecting and estimating depth information for real-time capture and generation of stereo video.

It is yet another object of the present invention to provide a monoscopic low-power mobile device that avoids computational complexity induced by using auxiliary equipment or human-interaction as well as SfM or depth analysis.

It is yet another object of the present invention to provide a monoscopic low-power mobile device that utilizes a low-complexity stereo image pair generation process.

It is another object of the present invention to provide a monoscopic low-power mobile device for capturing and generating stereo images with good 3D effects.

It is another object of the present invention to provide a monoscopic low-power mobile device that can be used for both stereo image and video capture and generation.

The objects are operable to capture an image, and operable to determine the best focus position by moving the lens over the entire focusing range through the focusing process and to select the focus position with the maximum focus value when capturing the image. Performed by a monoscopic low-power mobile device comprising a single-sensor camera sensor module with an autofocusing sub-module. The device includes a depth map generator assembly operable to automatically deploy a block-level depth map using statistics from the autofocusing sub-module at a first stage and to deploy an image depth map at a second stage. The device also includes an image pair generator module operable to generate a missing second view from the captured image to produce 3D stereo left and right views.

The monoscopic low-power mobile device uses the autofocus function of the monoscopic camera sensor to estimate depth map information, which is the SfM of other proposed systems as well as auxiliary equipment or human-interaction used in other approaches. Or avoid computational complexity induced by using depth analysis.

Monoscopic low-power mobile devices with additional but optional motion estimation modules that improve the accuracy of depth map detection for stereo video generation can be used for both stereo image and video capture and generation.

Monoscopic low-power mobile devices use statistics from the autofocus process to detect and estimate depth information to produce stereo images. The use of the autofocus process is suitable for low-power devices due to the two-stage depth map estimation design. That is, at the first stage, a block-level depth map is detected using the autofocus process. The approximated image depth map is generated by using bilinear filtering at the second stage.

In addition, monoscopic low-power mobile devices use a low-complexity approach to detect and estimate depth information for real-time capture and generation of stereo video. The approach uses statistics from the motion estimate, autofocus processing, and historical data plus some heuristics to estimate the depth map.

Monoscopic low-power mobile devices use a low-complexity stereo image pair generation process by using Z-buffer based 3D surface reconstruction.

As another aspect of the invention, a method for generating a real-time stereo image with a monoscopic low-power mobile device includes the steps of capturing an image, autofocusing the lens, and moving the lens over the entire focusing range to achieve the best focus. Determining a position and selecting the focus position having the maximum focus value when capturing the image; Automatically generating a block-level depth map using statistics from the autofocusing step in a first stage and generating an image depth map in a second stage; And generating a missing second view from the captured image to produce 3D stereo left and right views.

As another aspect of the invention, a method of processing still images includes autofocusing a captured still image and estimating depth information of remote objects in the image to detect a block-level depth map; And approximating the image depth map from the block-level depth map.

Autofocusing processing includes processing an image using a coarse-to-fine depth detection process. In addition, the approximation step includes bilinearly filtering the block-level depth map to derive the approximated image depth map.

In another aspect, the present invention, when executed by a processor, determines a depth value of each focus block comprising corner points A, B, C, and D of the block-level depth map. , The following equation

And program code operable to bilinearly filter the image to determine the depth value (d _p ) of all the pixels in the block, wherein the corner points (A, B) of the block are Position values and depth values for, C and D) are (x _A , y _A , d _A ), (x _B , y _B , d _B ), (x _C , y _C , d _C ), (x _D , y _D , d _D ), and each pixel is represented by a point P (x _P , y _P , d _P ).

In another aspect of the invention, a still image capture device comprises: an autofocusing module operable to process captured still images and estimate depth information of remote objects in the image to detect a block-level depth map; An image depth map module operable to approximate an image depth map from a block-level depth map using bilinear filtering; And an image pair generator module operable to generate a missing second view from the captured image to produce a three-dimensional (3-D) stereo left and right view.

In another aspect of the invention, a video image capture device comprises: an autofocusing module operable to process a captured video clip and estimate depth information of remote objects in a scene; And a video coding module operable to code the captured video clip, provide statistical information, and determine motion estimation. The depth map generator assembly detects and estimates depth information for real-time capture and generation of stereo video using statistical information from motion estimation, the process of the autofocusing module, and historical data plus heuristics, to obtain an image. Is operable to obtain a final block depth map from which the depth map is derived.

Brief description of the drawings

In addition to the foregoing summary, the following detailed description of the preferred embodiments of the present invention will be more readily understood when read in conjunction with the accompanying drawings. To illustrate the invention, presently preferred embodiments are shown in the drawings. However, it should be understood that the invention is not limited to the exact arrangement of the processes depicted.

1 shows a general block diagram of a monoscopic low-power mobile device.

2 illustrates a general block diagram of operations for capturing, processing, and displaying both real-time stereo image and video data.

3 shows a general block diagram of the operation of real-time capture and generation of a 3D still image.

4 shows a diagram of the relationship between lens position and object distance from a focus point.

5A shows a graph of the relationship between FV and lens position using a global search algorithm.

5B shows a graph of the relationship between FV and lens position for a coarse-to-fine search algorithm.

6A shows the original image.

6B shows an image depth map of the image of FIG. 6A.

FIG. 6C shows a block depth map of the image of FIG. 6A.

FIG. 6D shows a synthesized 3D anaglyph view using the block depth map of FIG. 6C.

6E shows a filtered depth map of the image of FIG. 6B.

7A shows a diagram of an intermediate point with neighboring blocks.

7B shows a diagram of a block with corner points.

8 shows a flow chart for a depth map generation process.

9A and 9B show the image of the first frame and the corresponding BDM.

9C and 9D show the thirtieth frame of the video and its corresponding BDM.

9E and 9F show the 60th frame of video and its corresponding BDM.

10A, 10B and 10C show an image depth map (IDM) generated from the BDMs shown in FIGS. 9B, 9D and 9F.

11 shows an image pair generation process.

12A shows a left view and a right view of binocular vision.

12B shows a geometric model of binocular vision for parameters for calculating a disparity map.

FIG. 13A shows an anaglyph image generated by using the approximated image depth map shown in FIG. 6E.

FIG. 13B shows an anamorphic image generated by using the exact image depth map shown in FIG. 6B.

FIG. 14A illustrates an example of the resulting anaglyph video frame of FIG. 9A.

FIG. 14B shows an example of the resulting anaglyph video frame of FIG. 9C.

FIG. 14C shows an example of the resulting anaglyph video frame of FIG. 9E.

15A and 15B show a flow chart of a Z-buffer based 3D interpolation process.

Detailed Description of the Preferred Embodiments

While the invention allows for embodiments in many different forms, these descriptions and the accompanying drawings disclose only a few forms as examples of embodiments of the invention. It is intended that the invention not be limited to the described embodiments, but the scope of the invention will be set forth in the appended claims.

Preferred embodiments of devices for capturing and generating stereo images and video in accordance with the present invention are described below for specific applications for monoscopic low-power mobile devices such as hand-held cameras, camcorders, or single-sensor camera phones. However, one skilled in the art will recognize that the present invention is also well suited for other types of devices with single-sensor camera modules. Next, referring to the drawings in which like reference numerals are used to denote like elements throughout the specification, a monoscopic low-power mobile device according to the present invention, generally designated 10, is shown in FIG. 1. .

In general, monoscopic low-power mobile device 10 includes a processor 56, a lens 12, and a single-sensor camera unit, hand-held digital, for controlling the operation of device 10 described herein. Camera sensor module 14, such as a camera or a camcorder. Processor 56 executes program instructions or programming code stored in memory 60 to perform the operations described herein. Storage 62 is a file system of a camera, camcorder, or single-sensor unit, and may include flash, disk, or tape depending on the application.

Camera sensor module 14 is capable of capturing still images in still image mode 18, and capable of capturing videos over a recording period in video mode 20 to produce a video clip. And (16). The camera sensor module 14 also includes an autofocusing sub-module 22 having dual modes of operation, namely still image mode 24 and video mode 26.

In addition, the monoscopic low-power mobile device 10 includes a depth map detector module 28 which also has a dual mode of operation, ie still image mode 30 and video mode 32. In the exemplary embodiment, depth map generator assembly 34 uses a two-stage depth map estimation process with dual modes of operation. As very well observed in FIGS. 2 and 3, the first stage (stage 1) of the two-stage depth map estimation process is performed in the auto mode of the still mode 24 performed by the autofocusing sub-module 22. Automatically develop the block-level depth map using the statistics from focusing processing 124 or autofocusing processing 126 in video mode 26. In a second stage, an image depth map is generated by the depth detection process 130 of the still mode 30 or the depth detection process 132 of the video mode 32 performed by the depth map detector module 28. In FIG. 2, f _i represents the i-th frame, f _i-1 represents the i-1th frame, d _i represents the block depth map (BDM) of the i-th frame, and d _i 'represents the i-th frame. Represents an image depth map (IDM).

The monoscopic low-power mobile device 10 has a single-sensor camera sensor module 14. Thus, only one image is captured and such image is used to represent the left (L) view for stereo imaging and display. An image pair generator module 42 is included in the device 10 to generate a second or missing right (R) view from the left view (original captured image) and the image depth map in the stereo view generator sub-module 48. do. Image pair generator module 42 also includes a disparity map sub-module 44 and a Z-buffer 3D surface reconstruction sub-module 46.

In an exemplary embodiment, 3D effects are displayed on display 58 using 3D effect generator module 52. In the exemplary embodiment, the 3D effect generator module 52 is an inexpensive red-blue anagrip to demonstrate the resulting 3D effect. The generated stereo views are easily displayed by other mechanisms such as holographic and stereoscopic devices.

Optionally, monoscopic low-power mobile device 10 includes a video coding module 54 for use in coding video. Video coding module 54 provides motion (estimation) information 36 for use in depth detection process 132 of video mode 32 by depth map detector module 28.

Referring also to FIG. 3, in operation, camera sensor module 14 captures one or more still images in still image mode 18 in image capture sub-module 16. Still image mode 18 performs capture process 118. Acquisition process 118 precedes autofocusing processing 124. In general, autofocusing processing 124 of still image mode 24 is used to estimate depth information of remote objects in a scene. To reduce computational complexity, block depth detection in stage 1 utilizes a coarse-to-fine depth detection algorithm in the exhaustive focusing search 125 of still image mode 24. The coarse-to-fine depth detection algorithm splits the image captured by capture process 118 in still image mode 18 into a number of blocks that detect the relevant depth map at the previous stage (stage 1). In stage 2, the depth detection process 130 of still image mode 30 uses bilinear filter 131B to derive an approximated image depth map from the block depth map of stage 1.

The autofocusing sub-module 22 of the still image mode 24 uses the global search focusing 125 used for still-image capture. In order to achieve real-time capture of video clips into video image mode 26, global search focusing 125 used for still-image capture is replaced with climb-hill focusing 127, Depth detection process 132 of video sub-module 32 includes motion information 36 from video coding module 54, focus value 38B from autofocusing process 126, and shown in FIG. Based on the frame history statistics 40, the block depth map 34 is detected.

Auto Depth Map Detection

Referring also to FIG. 3, the monoscopic low-power mobile device 10 uses the autofocusing process 124 of the autofocusing sub-module 22 for automatic block depth map detection. In image capture in still-image mode 18 and video mode 20 of operation, different approaches are needed due to the different focal length search algorithms used in this scenario (modes of operation).

In digital cameras, most focusing assemblies select the best focus position by evaluating image contrast on the imager plane. The focus value (FV) 38B is a score measured via the focus metric over a particular region of interest, and in general, the autofocusing process 126 selects the position corresponding to the highest focus value as the best focus position of the lens 12. do. In some cameras, the high frequency content of the image is used as the focus value (FV; 38B), for example

A high pass filter (HPF) such as can be used to capture high frequency components to determine the focus value (FV) 38B. Also, the focus value FV is an FV map as will be described later in the video mode.

There is a relationship between the lens position of the lens 12 from the focus point FV 38B (as shown in FIG. 4) and the target distance from the camera or device 10 with the camera, the relationship being specific It is fixed relative to the camera sensor module 14. Various camera sensors may have different statistics of such a relationship. Thus, once the autofocusing process 124 of the autofocusing sub-module 22 positions the best focus position of the lens 12 based on the information of the characteristics of the camera sensor module, the target object and the camera or device 10 The actual distance is estimated between), which is also the depth of the object in the scene. Thus, the depth map detection process depends on the sensor-dependent autofocusing process 124 or 126.

In still-image capture mode 18, most of the digital camera sensor modules 14 select a global search algorithm 125 for the autofocusing process 124, which is adapted to focus the lens 12 of the module. The best focus position is determined by selecting a focus position that moves across the entire focusing range and has the maximum focus value.

A typical example of global search algorithm 125 is the global search described with respect to FIG. 5A, which scans the entire focus range with the minimum motor step indicated by the rows of evenly-spaced arrow heads. On the other hand, FIG. 5B shows first the entire focus range using the larger step indicated by the row of arrow heads, and then the smaller step indicated by the arrow heads with the smaller distance between adjacent arrow heads. Use to illustrate a coarse-to-fine search that searches around the peak position.

Clearly, the accuracy of the depth map generated for a still-image depends entirely on the size of the spot focus windows selected for that image. In general, in the autofocusing process 124 for still-image mode 24, the image is split into N × N subblocks, also referred to as spot focus windows, and the focus value 38B is an autofocusing process. 124 is calculated for each focus window.

After the global search 125, the best focus position of the lens 12 is obtained for each focus window, so that the depth of the object corresponding to each window can be estimated. Clearly, minimal focus window size, better accuracy of the depth map, and higher computational complexity are achieved.

In monoscopic low-power mobile device 10, two types of depth maps, namely image depth map (IDM) and block depth map (BDM), are defined in depth map generator assembly 34. The pixel depth values of all the pixels in the image depth map are stored by the depth detection process 130, and the depth values of each focus window in the block depth map. In FIG. 6B, the image depth map 75 corresponding to the still-image 70 shown in FIG. 6A is obtained by setting the focus window size as 1 × 1, so that the image depth map 75 is pixel-level Presence at accuracy, where pixels with higher intensity correspond to objects closer to the viewpoint. In general, however, this setting is not possible for most applications due to excessive computational complexity for auto focusing. An example of the block depth map 77 is shown in FIG. 6C, where N is set to 11, which is a more practical setting for cameras with general computing power.

In general, the block depth map 77 generated in stage 1 by the autofocusing process 124 needs to be further processed to obtain an image depth map 80 (FIG. 6E), otherwise some artifacts may be May appear. An example of a synthesized 3D anaglyph view 79 using the block depth map 77 shown in FIG. 6C is shown in FIG. 6D, where a sharp depth gap between neighboring focus windows at the edges is imaged. Artifacts appear due to the fact that they do not correspond to actual object shape boundaries. The artifacts can be reduced by an artifact reduction process 131A that precedes the processing by the bilinear filter 131B. The filtered image depth map 80 is shown in FIG. 6E.

As best shown in FIGS. 7A and 7B, the artifact reduction process 131A consists of two steps. In a first step, the depth values of the corner points A, B, C and D of each block of FIG. 6C are found during the autofocusing process 124, and the depth values thereof are as shown in FIG. 7A. Will be the average of the neighboring blocks, where the depth d of the intermediate point is

Is defined by Equation (1), where d1, d2, d3 and d4 are depth values of neighboring blocks.

The block depth map generated by the autofocusing process 124 includes the depth value of each focus window / block stored. In FIG. 3, memory 60, which is a hardware block, and / or storage 62 (shown in FIG. 2) is not shown in the illustrated process.

After the depth values of all corner points A, B, C and D have been obtained, in the second step as best shown in FIG. 7B, the bilinear filtering obtains the depth values of the pixels inside the blocks. . As an example shown in FIG. 7B, the position and depth values for the corner points A, B, C and D of the block are (x _A , y _A , d _A ), (x _B , y _B , d _B ), (x _C , y _C , d _C ), (x _D , y _D , d _D ), so that the depth values of all the pixels in the block can be calculated. For example, for the pixel indicated by point P (x _P , y _P , d _P ), the pixel depth value d _p is

It can be obtained by the following equation (2).

Next, referring to FIG. 2, for video, global search algorithm 125 is not possible due to excessive delay caused in determining the best focus. Hill climbing focusing 127 is more general because of its faster seek speed. It seeks the best focus position as if climbing the heel. When camera sensor module 14 starts recording video in video mode 20 in image capture sub-module 16, global search algorithm 125 is used to find the best focus position as the initial position, After the initial lens position has been positioned, the camera sensor module 14 needs to determine in real-time how the focus lens 12 should move and how much it must move to reach the top of the heel. Clearly, during video mode 26 of autofocusing process 126, obtaining an accurate depth map for videos is much more difficult than for still-images. While not wishing to be bound by theory, this is because hill climbing focusing alone does not guarantee depth accuracy for other blocks, but obtains the correct depth for the area in focus. In addition, since the global search algorithm 125, which guarantees the accuracy of the depth for all blocks, is called only at the beginning of the recording, the image acquisition sub-module 16 applies to all blocks during the recording period of the video mode 20. It is impossible to correct the depth of

Next, referring to FIG. 8, a flowchart of a depth map detection process 132 for use in video mode 32 by depth map detector module 28 is shown. The current frame index is represented by _n , where {D _n (i, j)} and {F _n (i, j)} (i = 1,2, ... N, j = 1,2, .. .N) is the final determined block depth map (BDM) and focus value (FV) map 38A of the current frame, and {M _n (i, j)} and {V _n (i, j)} are autofocusing An internal BDM and FV map obtained by process 126, where {P _n (i, j)} and {T _n (i, j)} are internal BDM and FV maps obtained by motion prediction.

During the depth detection process 132 of the video mode 32, first, in step S134, it is determined by the focus position rk hill climbing focusing 127 of the current frame n, and the corresponding block depth map {M _n (i, j) } And FV map 38B {V _n (i, j)} are obtained. Step S134 is followed by step S136 where a determination is made as to whether motion information (MV) 36 is available from the video coding process 154 performed by the video coding module 54. If the determination is YES, the motion information (MV) 36 is analyzed and the global motion vector (GMV) is obtained in step S138. Step S138 is followed by step S139 where a determination is made as to whether the global motion (ie, GMV) is greater than the threshold. If the determination is YES, the lens 12 moves to other scenes, and then the tasks of maintaining the correct scene depth history and estimating the direction of object movement use a different process.

If the determination in step S139 is YES, then D _n (i, j) = M _n (i, j) and F _n (i, j) = V _n (i, j), and in step S144, Clean up the stored BDM and FV map history of previous frames during the update process of the BDM and FV maps.

Returning to step S136 again, in some systems, motion information 36 is used due to all kinds of causes, for example, the video is not coded or the motion estimation module of the coding algorithm is turned off. not possible. Therefore, the determination at step S136 is "no", and step S136 is followed by step S144 to be described later. If the determination in step S136 is no, the process assumes that the motion vectors are zero for all blocks.

If motion information 36 is available, in step 139,

By equations (3) and (4), the process 132 allows the BDM and FV of the current frames P _n (i, j) and T _n (i, j) from the BDM and FV maps of the previous frame. Predicting the map is followed by step 142, where blocks (a, b) of the (n-1) th frame are predictions of blocks (i, j) of the n th frame, and FV_TH is a threshold for FV differences. .

Step S142 is followed by step S144 where the device 10 assumes that the good focus delivers a more accurate depth estimate. Thus, the focus lens position corresponds to the maximum FV and is treated as the best choice. Based on that logic, the final BDM and FV maps are

Is determined by equations (5) and (6), where {D _n (i, j)} and {F _n (i, j)} (i = 1,2, ... N , j = 1,2, ..., N are the final determined block depth map (BDM) and focus value (FV) map 38A of the current frame; {M _n (i, j)} and {V _n (i, j)} are internal BDM and FV maps obtained by the autofocusing process 126; {P _n (i, j)} and {T _n (i, j)} are internal BDM and FV maps obtained by motion prediction.

As expected, equations (5) and (6) are not accurate for all cases. Equations (5) and (6) will fail for some difficult scenarios, such as when occlusion / exposure occurs. In general, it is reasonable to assume that video frames are captured at a speed of 15 to 30 frames per second, and since the objects in those frames are moving at reasonable speeds, the objects will not move too far away from neighboring frames.

Heuristic laws refer to the assumptions and logic for the above equations (3) to (6), and the assumptions and logic in the flowchart shown in FIG. These laws are actually coded in the program, and historical data is stored in memory 60 to estimate the depth map.

After the BDM is obtained, based on the same approach described with respect to the depth detection process 130 for the still image mode, an image depth map (IDM) is calculated in step S146 from the BDM results of step S144. Therefore, the BDM of step S144 is affected by the bilinear filtering by the artifact reduction 131A and the bilinear filter 131B (Fig. 3).

Returning to step S139, if the determination is no, step S139 is followed by step S140 in which the history rolls over. In step S140, the rollover history refers to the following actions, i.e., if a global motion (i.e. GMV is greater than the threshold) is detected, which means that the camera lens is moving to other scenes, then the correct scene history is obtained. The task of holding and estimating the direction of object movement becomes difficult. In this case, set D _n (i, j) = M _n (i, j) and F _n (i, j) = V _n (i, j), and store the stored BDM and FV map history of the previous frames. Clean up. Thereafter, step S146 is followed by step S146.

An example for demonstrating the process of FIG. 8 is shown in FIGS. 9A-9F. 9A and 9B show an image of the first frame 82 and the corresponding BDM 84. 9C and 9D, on the other hand, show the thirtieth frame 86 of the video and its corresponding BDM 88. FIG. 9E and 9F show the 60th frame 90 of the video and its corresponding BDM 92. In the video, the plastic bottle rolls from a distance to the camera. It can be readily observed from these figures that the process 132 can catch the movement of objects in the scene and reflect these activities in the obtained depth map.

10A, 10B, and 10C, image depth maps IDM 94, 96, and 98 are the BDMs 84, 88, and 9 shown in FIGS. 9B, 9D, and 9F, respectively, using process 132. 92). IDMs 94, 96, and 98 are obtained by using depth detection process 130 (FIG. 3).

Create stereo image pair

1 and 11, so far, device 10 has captured an image or left view and acquired a corresponding image depth map. Image pair generation module 42 uses an image pair generation process 142, which will be described next. In step S144 the left view is obtained, and in step S146 its corresponding image depth map is obtained from the depth detection process 130 or 132.

First, although the image pair generation process 142 assumes that the acquired image is a left view of the stereoscopic system in step S144, the image may alternatively be considered a right view. Then, based on the image depth map obtained in step S146, in step S148, a disparity map (pixel distance between image points in both views) for that image is obtained in disparity map sub-module 44. Is calculated. Disparity map calculations by the disparity map sub-module 44 will be described below with reference to FIGS. 12A and 12B. However, both the left view and the depth map are also inputs for calculating the disparity map for 3D view generation, and the left view and depth map contribute directly to the Z-buffer based surface reconstruction. Step S148 is followed by step S150 where the Z-buffer based 3D interpolation process 146 by the Z-buffer 3D surface reconstruction sub-module 46 is called to construct a 3D visible surface for the scene from the right eye. Step S150 is followed by step S152 in which the right view is obtained by projecting the 3D surface onto the projection plane, as best shown in FIG. 12B. Step S152 is performed by the stereo view generator sub-module 48.

In FIG. 12A, using the left (L) view and the right (R) view on the projection plane for the far object, a geometric model of binocular vision is shown. In FIG. 12B, F is the focal length, L (x _L , y _L , 0) is the left eye, R (x _R , y _R , 0) is the right eye, and T (x _T , y _T , z) Are the 3D points in the scene, and P (x _P , y _P , F) and Q (x _Q , y _Q , F) are projection points of T on the left and right projection planes. Clearly, the horizontal positions of P and Q on the projection planes are (x _P -x _L ) and (x _Q -x _R ), so the disparity is d = [(x _Q -x _R )-(x _P -x _L )].

As shown in Figure 12b, the ratio of F and z,

Is defined in equation (7), where z is depth. Therefore, equations (8) and (9) are

And thus the disparity d is

It can be obtained by the following equation (10).

Thus, for all the pixels in the left view, its counterpart in the right view is shifted to the left side or right side by the distance of the disparity value obtained in equation (10). However, the mapping from the left-view to the right-view is not a 1-to-1 mapping due to possible occlusions and therefore requires additional processing to obtain the right-view image.

Thus, Z-buffer based 3D interpolation process 146 is performed by Z-buffer 3D surface reconstruction sub-module 46 for right-view generation. Since the distance between the two eyes is very small compared to the distance from the eye to the object (as shown in FIG. 12A), approximately the distance from the object to the left eye is equal to the distance from the object to the right eye. Consider, this will simplify the calculation very much. Thus, the depth map Z (x, y), where Z (x, y) is the actual image depth map, but is an unknown map to be detected, is retained for the right (R) view, where x, y is Pixel position in the view.

Next, referring to FIGS. 15A and 15B, a process 146 for reconstructing the 3D visible surface for the right-view will next be described. At the beginning (step S166), the depth map is initialized as infinity. Step S166 follows step 168 in which pixels (x ₀ , y ₀ ) of the left view are obtained. Then, in step S170, for every pixel (x ₀ , y ₀ ) of the left view having a depth z ₀ and a disparity value d ₀ ,

The depth map is updated for its corresponding pixel in the right view by equation (11) defined as

Step S170 is followed by step S172, which is a determination step of determining whether any more pixels are present. If the determination is YES, step S172 returns to step S168 to obtain the next pixel. On the other hand, after all the pixels of the left-view have been processed and accordingly the determination in step S172 is "no", in step S172, the reconstructed depth map is checked and pixels having values that are infinite (valid on the left-view). Step S174, which is searched for pixels without a map, is followed. Step S174 is followed by step S176 in which a determination is made as to whether the pixel value PV is infinity. If the determination in step S176 is "no", the pixel value PV is valid and can be used directly as the intensity value in step S188 of Fig. 15B.

If the determination in step S176 is YES, then in step S180, for such pixels, the depth for the corresponding pixel is first calculated by 2D interpolation based on its neighboring pixels with available depth values. Thereafter, the disparity value is calculated using the above equation (10) in step S182, and then, in step S184, the corresponding pixel of the pixel is searched backward in the left view. Step S184 is followed by step S186 of determining whether or not a pixel has been found. If the corresponding pixel is available, step S186 is followed by step S188, where a corresponding intensity value can be used for the right-view pixel. Otherwise, if the determination in step S186 is no, then step S186 is followed by step S190 of using interpolation to calculate the intensity value based on neighboring pixels in the right-view with the available intensity values. .

It is important to indicate that the advantage of using the proposed algorithm over the direct intensity interpolation method is that it takes into account the 3D continuity of the object shape resulting in better realism for stereo effects. Clearly, the problem of restoring the invisible area of the left view is an unsolved problem. In one known solution, the depth of a missing pixel is reconstructed by using its neighboring pixels in the horizontal direction corresponding to the additional surface, assuming that there are no other visible surfaces behind the scene. In some cases, the assumption may not be valid. In order to consider more possible cases, in the proposed solution, surface reconstruction takes into account the depth of all neighboring pixels in all directions, which will reduce the chance of invalid assumptions and better 3D continuity of the reconstructed surface. Will effect.

Experiment result

Device 10 may be implemented in an MSM8K VFE C-SIM system. Experimental results indicate that the captured and generated stereo images and videos have a good 3D effect.

In the experiments, cheap red-blue anaglyph generation process 152 was used to demonstrate the resulting 3D effect, but it is possible that the generated stereo views are displayed by other mechanisms such as holographic and stereoscopic devices. In the first experiment, as shown in Figs. 13A and 13B, stereo image pairs were calculated using different kinds of image depth maps and generated corresponding anaglyph images. FIG. 13A is generated by using the approximated image depth map shown in FIG. 6E, and FIG. 13B is generated by using the exact image depth map shown in FIG. 6B. Clearly, the results indicate that the approximated image depth map results in image quality similar to using an accurate depth map, which demonstrates good performance.

In summary, monoscopic low-power mobile device 10 provides real-time capture and generation of stereo images and video. The device 10 uses the autofocusing process of the monoscopic camera sensor module 14 to capture and generate stereo images and video. The autofocusing process of the camera sensor is used to estimate depth information of remote objects in the scene. In video capture, a low-complexity algorithm is provided to detect a block depth map based on motion information, focus values, and frame history statistics.

Device 10 is configured for real-time applications such that computational complexity is of primary concern. However, device 10 estimates the object depth in a coarse-to-fine manner, that is, the image is divided into a number of blocks so that the relevant block depth map can be detected quickly. A bilinear filter is then used to transform the block depth map into an approximated image depth map. In stereo image generation, a low-complexity Z-buffer based 3D surface reconstruction approach is used to estimate missing views.

Experimental results indicate that the captured and generated stereo images and video have sufficient 3D effects. The better the focus function of the sensor module 14, the more accurate the estimated depth map is, and therefore the better the stereo effect the generated image and video has.

The foregoing description of the embodiments of the invention has been presented for the purposes of illustration and description. It is not intended to be exhaustive or to limit the invention to the precise form disclosed, and modifications and variations are possible in light of the above teachings or may be acquired from practice of the invention. Embodiments have been selected and described in order to explain the principles of the invention and its practical applications in order that the present invention may be utilized by those skilled in the art in various embodiments and in various variations suitable for the particular use contemplated. It is intended that the scope of the invention be defined by the claims appended hereto and their equivalents.

Claims (44)

An autofocusing sub-module operable to capture an image and operable to determine the focus position by moving the lens over the entire focusing range through a focusing process and selecting the focus position with the maximum focus value when capturing the image Single-sensor camera sensor module having a; A depth map operable to automatically develop a block-level depth map using statistics from the autofocusing sub-module in a first stage and an image depth map using the block-level depth map in a second stage Generator assembly; And An image pair generator module operable to generate a missing second view from the captured image to generate a three-dimensional (3D) stereo left and right view using the image depth map; The focusing process of the autofocusing sub-module performs an exhaustive search focusing process to capture still-images in still image mode, and to achieve real-time capture of a video clip in video mode. A monoscopic low-power mobile device initiated by a search focusing process, followed by a climbing-hill focusing process. The method of claim 1, The image pair generator module, A disparity map sub-module that calculates a disparity map based on pixel distances between image points in a left view and a right view of a binocular vision geometry for the captured image, The disparity map sub-module, wherein the captured image represents the left view; A Z-buffer 3D surface reconstruction sub-module operable to construct a 3D visible surface for the captured image from the right view; And And a stereo view generator sub-module operable to project the 3D surface of the right view into a projection plane. delete The method of claim 1, In the second stage, the depth map generator assembly is operable to reduce artifacts with a bilinear filter. The method of claim 4, wherein In the second stage, the depth map generator assembly is operable to obtain a depth value of corner points (A, B, C, and D) of each block found during the focusing process, the depth value being intermediate The depth d of the point

An average value of neighboring blocks of each block, defined by d ₁ , d ₂ , d _3, and d ₄ are depth values of the neighboring blocks; After the depth value of all of the corner points (A, B, C and D) has been obtained, the bilinear filter is operable to obtain a pixel depth value of the pixels in the blocks. The method of claim 5, The bilinear filter is also,

Is operable to calculate a pixel depth value d _p for pixel P (x _P , y _P , d _P ) according to the equation Position values and depth values for the corner points A, B, C and D of each block are (x _A , y _A , d _A ), (x _B , y _B , d _B ), ( monoscopic low-power mobile device, represented as x _C , y _C , d _C ), (x _D , y _D , d _D ). The method of claim 1, A video coding module for coding the captured video clip and providing statistical information to calculate the block-level depth map, The video coding module is operable to determine a motion estimate, The depth map generator assembly may, in the second stage, obtain statistical information from the motion estimate, the focusing process, and historical data plus heuristic law to obtain a final block depth map from which the image depth map is derived. A monoscopic low-power mobile device operable to detect and estimate depth information for real-time capture and generation of stereo video using data plus heuristic rules. The method of claim 1, And a 3D effect generator module for displaying the 3D stereo left and right views on the display. The method of claim 8, And the 3D effect generator module is operable to generate a red-blue anaglyph image of 3D stereo left and right views on the display. The method of claim 1, A monoscopic low-power mobile device comprising one of a hand-held digital camera, a camcorder, and a single-sensor camera phone. Means for capturing an image with a single sensor; Autofocusing means for autofocusing a lens and determining a focus position by moving the lens over the entire focusing range and selecting a focus position having a maximum focus value when capturing the image; Means for automatically generating a block-level depth map using statistics from the autofocusing means in a first stage and generating an image depth map using the block-level depth map in a second stage; And Means for generating a missing second view from the captured image to create a three-dimensional (3D) stereo left and right view using the image depth map, The auto focusing means, Means for performing a global search focusing process to capture still-images in still image mode; Means for initiating the global search focusing process in a video mode; And And means for climbing-hill focusing in video mode to capture real-time video clips. The method of claim 11, Means for generating the missing second view, Means for calculating a disparity map based on pixel distances between image points in a left and right view of a binocular vision geometry for the captured image, wherein the captured image represents the left view; Means for calculating; 3D surface reconstruction means using Z-buffering, constituting a 3D visible surface for the captured image from the missing right viewpoint; And And means for generating a stereo view by projecting the constructed 3D surface into a projection plane. delete The method of claim 11, And the means for generating the image depth map at the second stage comprises means for reducing artifacts. The method of claim 14, Means for reducing the artifacts, Means for obtaining a depth value of corner points (A, B, C, and D) of each block found by said autofocusing means, wherein the depth value is equal to the depth d of the intermediate point.

Means for obtaining the depth value, wherein an average value of neighboring blocks of each block is defined by d ₁ , d ₂ , d _3, and d ₄ are depth values of the neighboring blocks; And Means for bilinear filtering to obtain a pixel depth value of the pixels in the blocks after the depth value of all of the corner points (A, B, C and D) is obtained. . The method of claim 15, The means for filtering bilinearly,

Means for calculating a pixel depth value d _p for pixel P (x _P , y _P , d _P ) according to the equation Position values and depth values for the corner points A, B, C and D of each block are (x _A , y _A , d _A ), (x _B , y _B , d _B ), ( monoscopic low-power mobile device, represented as x _C , y _C , d _C ), (x _D , y _D , d _D ). The method of claim 11, Means for providing statistical information and video coding the captured video clip, The means for video coding comprises means for motion estimation, Means for generating the image depth map, Real-time capture of stereo video using statistical information from the means for motion estimation, the autofocusing means, and historical data plus heuristic laws, to obtain a final block depth map from which the image depth map is derived. Means for detecting and estimating depth information for generation. The method of claim 11, And a means for generating 3D effects of the 3D stereo left and right views on the display. The method of claim 18, And the 3D effect generating means generates a red-blue anaglyph image of the 3D stereo left and right views on the display. The method of claim 11, A monoscopic low-power mobile device comprising one of a hand-held digital camera, a camcorder, and a single-sensor camera phone. Capturing an image with a single sensor; Autofocusing a lens and determining a focus position by moving the lens over the entire focusing range and selecting a focus position having a maximum focus value when capturing the image; Automatically generating a block-level depth map using the statistics from the autofocusing step in a first stage, and generating an image depth map using the block-level depth map in a second stage; And Generating a missing second view from the captured image to produce a three-dimensional (3D) stereo left and right view using the image depth map, The autofocusing step, Performing a global search focusing process to capture still-images in still image mode; Initiating the global search focusing process in a video mode; And Climbing-hill focusing in video mode to capture real-time video clips. The method of claim 21, Generating the missing second view, Calculating a disparity map based on the pixel distance between the image points in the left and right views of the binocular vision geometry for the captured image, wherein the captured image represents the left view; Calculating; 3D surface reconstruction using Z-buffering to construct a 3D visible surface for the captured image from the missing right viewpoint; And Generating a missing right view by projecting the constructed 3D surface onto a projection plane. delete The method of claim 21, Generating the image depth map at the second stage comprises reducing artifacts. The method of claim 24, Reducing the artifacts, Obtaining a depth value of the corner points A, B, C, and D of each block found by the autofocusing step, wherein the depth value is equal to the depth d of the intermediate point.

Obtaining a depth value of an average value of neighboring blocks of each block, wherein d ₁ , d ₂ , d _3, and d ₄ are depth values of the neighboring blocks; And After the depth value of all the corner points (A, B, C and D) is obtained, performing bilinear filtering to obtain pixel depth values of the pixels in the blocks. How to produce. The method of claim 25, Performing the bilinear filtering,

Calculating a pixel depth value d _p for the pixel P (x _P , y _P , d _P ) according to the equation Position values and depth values for the corner points A, B, C and D of each block are (x _A , y _A , d _A ), (x _B , y _B , d _B ), ( x _C , y _C , d _C ), (x _D , y _D , d _D ). The method of claim 21, Video coding and motion estimation of the video clip, Generating the image depth map, Real-time capture and generation of stereo video using the statistics from the motion estimating step, the autofocusing step, and historical data plus heuristics to obtain a final block depth map from which the image depth map is derived. Detecting and estimating depth information for a real-time stereo image. The method of claim 21, Generating 3D effects of the 3D stereo left and right views on a display. 29. The method of claim 28, Generating the 3D effect comprises generating a red-blue anaglyph image of the 3D stereo left and right views on the display. delete delete delete delete delete delete delete delete delete delete delete delete delete delete delete

Images